WildActor keeps video characters consistent with just three photos

Anyone who has experimented with AI video knows the frustration: the first frame looks stunning, but the moment your character turns around, their face morphs into a stranger, their shirt changes color, or their limbs melt into the background.

Keeping characters consistent across wild camera angles and intense motions has long been one of the toughest nuts to crack. A research team from Meituan recently open-sourced WildActor, a project accepted to ICML 2026 designed specifically to solve this problem. Using just three reference photos, it keeps a character's face, body, and clothing locked in across unconstrained viewpoints, compositions, and movements.

Three photos make an unbreakable anchor

Most existing identity-preserving methods lean heavily on a single close-up face crop. That works fine until the character turns around or performs a full-body action, forcing the diffusion model to hallucinate what the back of their head or jacket looks like.

WildActor takes a much smarter route by utilizing canonical three-view references, covering front, side, and back views. By supplying these three anchor images, the model has ground truth references for virtually every camera angle, allowing the subject to sprint, jump, or spin without losing their identity.

Crucially, this setup is not limited to photorealistic humans. The pipeline works equally well on cartoon and stylized characters, where identity drift is often just as noticeable.

Built on Wan2.2-5B and open data pipelines

Rather than training a massive video base model from scratch, the team built WildActor on top of the open-source Wan2.2-5B backbone, integrating it cleanly with DiffSynth-Studio.

Inference is handled via lightweight adapter weights. Users configure a simple JSON file with their text prompt and image paths, point the script to the adapter checkpoint, and generate the final clip.

Alongside the inference code, the authors released the Actor-18M data construction pipeline. This toolkit filters identity-consistent single-person videos, pulls face and body references, and automatically produces normalized multi-view anchors for downstream training.

Why it matters

For digital creators and indie animators, consistent identity is the single biggest barrier between AI video being an interesting toy and a viable production pipeline. If you are building a serialized web show or game cinematics, you cannot afford to have your lead actor swap outfits every two seconds. WildActor turns character continuity into a plug-and-play step requiring only three basic turnaround shots.

For the broader industry, Meituan's approach shows that winning the character consistency game does not require pre-training another multi-billion-parameter foundation model from zero. By targeting multi-reference conditioning and open-sourcing the data filtering pipeline on top of Wan2.2-5B, they are giving the open-source ecosystem a modular way to handle dynamic character motion.

We think this lightweight adapter pattern is exactly what creators need right now, provided local hardware can comfortably handle 5B-scale video generation.